The study of computing in presence of faulty robots in the Look-Compute-Move model has been the object of extensive investigation, typically with the goal of designing algorithms tolerant to as many faults as possible. In this paper, we initiate a new line of investigation on the presence of faults, focusing on a rather different issue. We are interested in understanding the dynamics of a group of robots when they execute an algorithm designed for a fault-free environment, in presence of some undetectable crashed robots. We start this investigation focusing on the classic point-convergence algorithm by Ando et al. for robots with limited visibility, in a simple setting (which already presents serious challenges): the robots operate fully synchronously on a line, and at most two of them are faulty. Interestingly, and perhaps surprisingly, the presence of faults induces the robots to perform some form of scattering, rather than point-convergence. In fact, we discover that they arrange themselves inside the segment delimited by the two faults in interleaved sequences of equidistant robots. The structure that they form has a hierarchical nature: robots organize themselves in groups where a group of some level converges to an equidistant distribution only after all groups of lower levels have converged. This is the first study on the unintended dynamics of oblivious robots induced by the presence of faults.